CN116258551A - User characterization learning and recommending method and system - Google Patents
User characterization learning and recommending method and system Download PDFInfo
- Publication number
- CN116258551A CN116258551A CN202310135479.9A CN202310135479A CN116258551A CN 116258551 A CN116258551 A CN 116258551A CN 202310135479 A CN202310135479 A CN 202310135479A CN 116258551 A CN116258551 A CN 116258551A
- Authority
- CN
- China
- Prior art keywords
- user
- session
- characterization
- term
- representing
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000012512 characterization method Methods 0.000 title claims abstract description 133
- 238000000034 method Methods 0.000 title claims abstract description 34
- 230000007774 longterm Effects 0.000 claims abstract description 89
- 238000009826 distribution Methods 0.000 claims abstract description 35
- 238000013528 artificial neural network Methods 0.000 claims abstract description 11
- 230000003068 static effect Effects 0.000 claims description 47
- 230000006870 function Effects 0.000 claims description 35
- 239000011159 matrix material Substances 0.000 claims description 26
- 238000004364 calculation method Methods 0.000 claims description 23
- 238000011176 pooling Methods 0.000 claims description 19
- 238000006116 polymerization reaction Methods 0.000 claims description 10
- 230000002776 aggregation Effects 0.000 claims description 5
- 238000004220 aggregation Methods 0.000 claims description 5
- 238000005070 sampling Methods 0.000 claims description 4
- 238000012549 training Methods 0.000 claims description 4
- 230000004931 aggregating effect Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000006399 behavior Effects 0.000 description 1
- 230000003542 behavioural effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000001186 cumulative effect Effects 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000004806 packaging method and process Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Strategic Management (AREA)
- Marketing (AREA)
- Economics (AREA)
- Development Economics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- General Business, Economics & Management (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides a user characterization learning and recommending method and a system, comprising the following steps: using a fully connected neural network to convert one-hot codes of users and projects into an ebedding representation; dividing a user session into a long-term session and a short-term session, extracting key information, and respectively forming a long-term session characterization and a short-term session characterization; calculating probability distribution of a user group to which the user possibly belongs; according to probability distribution, the characteristics of different user groups are aggregated, the influence of the neighbors of the users and the preference difference between the subsets of the like-minded users are captured, and group influence characterization is obtained; constructing a hybrid user characterization by long-term session characterization and short-term session characterization and group influence characterization of the user; the probability of one item becoming the next access item is estimated. The invention solves the problem that in most of the existing session-based recommendation systems, the characterization of the user is summarized based on the session of the user independently, and the information sharing between user models is ignored.
Description
Technical Field
The invention relates to the technical field of recommendation systems, in particular to a user characterization learning and recommendation method and system, and especially relates to a user characterization learning and recommendation method and system based on long-term and short-term user grouping modeling.
Background
Session-based recommendation systems (SBRSs, session-based Recommender Systems) have become a current research hotspot and are increasingly being applied to the recommendation of the next item. In one session, a packaging scope of items is specified, such as a set of products in a shopping cart, a set of browsed websites within a certain time frame, etc. Different sessions reflect different preferences and needs of the user, as the interests of the user are changing over different time periods. Thus, session-based recommendations not only simulate the long-term preferences of the user, but also capture the short-term preferences of the user.
Most existing SBRS assume that the user's current preferences have different associations with long-term and short-term sessions, and they employ different networks to distinguish the different contributions of the long-term session to describe the user's current interests. These systems significantly improve the recommended performance compared to conventional approaches. However, in these methods, the characterization of the users is summarized independently based on their own session, so that the learned model is built on a per-user basis, with no explicit information sharing between the user's models.
However, in practical applications, dynamic groups of users who like the same channel exist in different contexts, and users in the same group typically have similar preferences. Therefore, if the group information is effectively utilized, more emphasis is placed on some related projects or users to construct a more targeted local recommendation model, the dynamic preference of the users and the continuously evolving potential group information can be better captured, and therefore the recommendation performance is improved.
The patent document with publication number of CN115204967A discloses a recommendation method which is integrated with implicit feedback of user long-term interest characterization, belongs to the field of recommendation system research in the field of machine learning, improves the implicit feedback part of a collaborative filtering method based on matrix decomposition, and not only considers the implicit feedback of users browsing or scoring commodities in a user commodity bipartite graph, but also considers the implicit feedback of similar-interest users. However, the patent document still has a defect that there is no clear information sharing between models of users.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to provide a user characterization learning and recommending method and system.
The invention provides a user characterization learning and recommending method, which comprises the following steps:
step 1: using a fully connected neural network to convert one-hot codes of users and projects into an ebedding representation;
step 2: dividing a user session into a long-term session and a short-term session, extracting key information based on the characterization of the items in the session obtained in the step 1, and respectively forming the long-term session characterization and the short-term session characterization through a pooling layer;
step 3: calculating probability distribution of a user group to which the user possibly belongs based on the long-term and short-term session characterization obtained in the step 2;
step 4: according to the probability distribution obtained in the step 3, the characteristics of different user groups are aggregated, the influence of the user neighbors and the preference difference between the aspiration channel user subsets are captured, and the group influence characterization is obtained;
step 5: constructing a hybrid user characterization by using the long-term session characterization and the short-term session characterization of the user obtained in the step 2 and the group influence characterization obtained in the step 4;
step 6: and (3) using the mixed user characterization obtained in the step 5 to replace the user potential vector in the paired model, and estimating the probability of one item to be the next access item.
Preferably, the step 1 specifically includes: one-hot transcoding of users and items using fully connected neural networksFor low-dimensional empdding characterization, a user matrix U E R is obtained N×K And an item matrix V E R M×K Wherein n= |u|, m= |v| represent the number of users and the number of items, respectively, K represents the implicit vector dimension, U e R K and v∈RK Representing the static representation of user u and the static representation of item v, respectively, R representing the matrix.
Preferably, the step 2 specifically includes: dividing a session of a user u into two parts, wherein the current sessionFor short-term sessions, the short-term needs of the user are described, and the rest sessions areFor long-term session, for representing long-term preference of user, wherein v i Representing a static representation of the ith item, +.>Representing the number of items in a short-term session, +.>Representing the number of items in the long-term session;
based on the item representations in each session, an item representation matrix is generated for the session and the item features are aggregated by a pooling operation to form two context-aware input representations x s and xl Reflecting the user's status without a time stamp.
Preferably, the specific calculation formula is as follows:
wherein ,xs and xl A short-term session characterization and a long-term session characterization respectively,representing a matrix spliced by static representations of items in a short-term session, < >>Representing a matrix stitched from static representations of items in a long-term session. The pooling represents a weighted average pooling operation, and the weight is calculated according to the static representation of the target user u and the static representation of the object v, and the specific calculation formula is as follows:
wherein ,wu,v Representing the weight generated for item v based on target user u, u representing the static characterization of user u, v representing the static characterization of item v,representing a long-term user session, v i Representing a static representation of the ith item in the long-term session, T represents a transpose operation, exp () represents an exponential function based on e.
Preferably, the step 3 specifically includes: long-short session characterization x generated based on step 2 s and xl Calculate them and potential user group G k Similarity between key representations of (a) and (b) evaluating probability distribution b of a group of users to which a user is currently likely to belong s u and bl u The specific calculation formula is as follows:
wherein ,bs u ,b l u ∈R L Representing probability distribution, x, of short-term and long-term user groups, respectively, to which a user is currently likely to belong s and xl Representing short-term session characterization and long-term session characterization, respectively, the softmax function will vector G k x s u And converting into a pseudo probability distribution vector.
Preferably, the step 4 specifically includes: according to the probability distribution b obtained in step 3 s u and bl u The characteristics of different user groups are aggregated, the influence of the neighbors of the users and the preference difference among the subsets of the like-minded users are captured, and the calculation formula of the characteristics of the aggregated user groups is as follows:
wherein ,gl u ∈R K ,g s u ∈R K Characterizing the influence of the long-term and short-term potential population of user u, G v Representing potential user groups, b s u and bl u Representing probability distributions for the short-term and long-term user groups, respectively, to which the user is currently likely to belong.
Preferably, the step 5 specifically includes: constructing a mixed user representation by utilizing the user long-short term session representation obtained in the step 2 and the influence representation of the current population and the history population obtained in the step 4, and carrying out { x } s u ,x l u ,g s u ,g l u Denoted as F.
Preferably, x is combined in a dynamic manner s u 、x l u 、g s u 、g l u Specific methods are using MLP polymerization or using attention polymerization;
polymerization using MLP: each feature vector is mapped to a scalar using a multi-layer perceptron, and then the scalar is converted to the weight of each component using a softmax layer, as specifically calculated below:
wherein F represents a mixed user representation, F represents a certain feature vector in F, w f Weights, h, generated for feature vector f u Representing the resulting hybrid user representation, MLP representing the multi-layer perceptron, T representing the transpose operation, exp () representing the e-based exponential function;
attention aggregation was used: the weight is calculated according to the relation between the component and the representation of the target user u, and the concrete calculation is as follows:
wherein F represents a mixed user representation, F represents a certain feature vector in F, u represents a static representation of user u, and w f Weights, h, generated for feature vector f u Representing the resulting hybrid user token, T represents the transpose operation and exp () represents the base e exponential function.
Preferably, the step 6 specifically includes: calculating step 5 to obtain the hybrid user representation h u The inner product of the token vector with the object v,taking the preference score of the user u on the article v as a predicted preference score, and calculating the following formula:
wherein ,representing a predicted preference score, h, of user u for item v u Representing a hybrid user characterization, v representing a static characterization of the item v, T representing a transpose operation;
training a model with ordering and pairwise penalty functions for positive samples v + Randomly selecting an item from the user's current session, for negative sampling v - Selecting a commodity that a user has never purchased or accessed before, the final loss function is as follows:
wherein , and />Representing the predicted user u versus the sample item v, respectively + And negatively sampled item v - Preference score of->Representing a long-term user session, σ representing an s igmoid function, θ representing parameters of the model, argmin representing a minimization function;
and selecting the item with the highest prediction score as the item which is most likely to be accessed next by the predicted user u by using the obtained trained model.
The invention also provides a user characterization learning and recommending system, which comprises the following steps: :
module M1: using a fully connected neural network to convert one-hot codes of users and projects into an ebedding representation;
module M2: dividing a user session into a long-term session and a short-term session, extracting key information based on the representation of the items in the session obtained in the module M1, and respectively forming the long-term session representation and the short-term session representation through a pooling layer;
module M3: based on the long-term and short-term session characterization obtained by the module M2, calculating probability distribution of a user group to which the user possibly belongs;
module M4: according to the probability distribution obtained by the module M3, the characteristics of different user groups are aggregated, the influence of the user neighbors and the preference difference between the aspiration channel user subsets are captured, and the group influence characterization is obtained;
module M5: constructing a hybrid user characterization by using the long-term session characterization and the short-term session characterization of the user obtained by the module M2 and the group influence characterization obtained by the module M4;
module M6: the mixed user token obtained by module M5 is used to estimate the probability of one item becoming the next access item instead of the user potential vector in the paired model.
Compared with the prior art, the invention has the following beneficial effects:
1. the invention solves the problem that in most of the existing session-based recommendation systems, the characterization of the user is summarized based on the session of the user independently, and the information sharing between user models is ignored;
2. when recommending articles for users, the method comprehensively considers the influence of the dynamic interest information and potential groups of the users, and can help merchants to capture user preferences with richer and more accurate contained information, so that the recommending accuracy and the adhesiveness of the users are improved, the demands of the users are better met, and higher economic benefits are brought to the merchants;
3. the invention can capture the potential group information of the user and sense the change condition of the group information of the user along with time.
Drawings
Other features, objects and advantages of the present invention will become more apparent upon reading of the detailed description of non-limiting embodiments, given with reference to the accompanying drawings in which:
FIG. 1 is a flow chart of the steps of the user characterization learning and recommendation method of the present invention.
Detailed Description
The present invention will be described in detail with reference to specific examples. The following examples will assist those skilled in the art in further understanding the present invention, but are not intended to limit the invention in any way. It should be noted that variations and modifications could be made by those skilled in the art without departing from the inventive concept. These are all within the scope of the present invention.
Example 1:
as shown in fig. 1, the present embodiment provides a user characterization learning and recommending method, which includes the following steps:
step 1: using a fully connected neural network to convert one-hot codes of users and projects into an ebedding representation; the step 1 specifically comprises the following steps: using a fully connected neural network to convert one-hot codes of users and articles into ebedding representation to obtain a user matrix U E R N×K And an item matrix V E R M×K Wherein n= |u|, m= |v| represent the number of users and the number of items, respectively, K represents the implicit vector dimension, U e R K and v∈RK Representing the static representation of user u and the static representation of item v, respectively, R representing the matrix.
Step 2: dividing a user session into a long-term session and a short-term session, extracting key information based on the characterization of the items in the session obtained in the step 1, and respectively forming the long-term session characterization and the short-term session characterization through a pooling layer; dividing a session of a user u into two parts, wherein the current sessionFor short-term session, the user's short-term needs are described, the remaining sessions are +.>For long-term session, for representing long-term preference of user, wherein v i Representing a static representation of the ith item, +.>Representing the number of items in a short-term session, +.>Representing the number of items in the long-term session;
based on the item representations in each session, an item representation matrix is generated for the session and the item features are aggregated by a pooling operation to form two context-aware input representations x s and xl Reflecting the state of the user without the time stamp;
the specific calculation formula is as follows:
wherein ,xs and xl A short-term session characterization and a long-term session characterization respectively,representing a matrix spliced by static representations of items in a short-term session, < >>Representing a matrix stitched from static representations of items in a long-term session. pooling represents a weighted average pooling operation. According to the static representation of the target user u and the static representation of the object v, calculating the weight according to the specific calculation formula as follows:
wherein ,wu,v Representing the weight generated for item v based on target user u, u representing the static characterization of user u, v representing the static characterization of item v,representing a long-term user session, v i Representing a static characterization of the ith item in the long-term session. T denotes a transpose operation, exp () denotes an exponential function based on e.
Step 3: based on the long-short session characterization obtained in the step 2, calculating probability distribution of a user group to which the user possibly belongs; the step 3 is specifically as follows: long-short session characterization x generated based on step 2 s and xl Calculate them and potential user group G k Similarity between key representations of (a) and (b) evaluating probability distribution b of a group of users to which a user is currently likely to belong s u and bs u The specific calculation formula is as follows:
wherein ,bs u ,b l u ∈R L Representing probability distribution, x, of short-term and long-term user groups, respectively, to which a user is currently likely to belong s and xl Representing short-term session characterization and long-term session characterization, respectively. The softmax function will vector G k x s u And converting into a pseudo probability distribution vector.
Step 4: according to the probability distribution obtained in the step 3, the characteristics of different user groups are aggregated, and the capturing is carried outThe influence of the user neighbors and the preference difference between the subsets of the like-minded users obtain group influence characterization; the step 4 is specifically as follows: according to the probability distribution b obtained in step 3 s u and bs u The characteristics of different user groups are aggregated, the influence of the neighbors of the users and the preference difference among the subsets of the like-minded users are captured, and the calculation formula of the characteristics of the aggregated user groups is as follows:
wherein ,gl u ∈R K ,g s u ∈R K The impact of the long-term and short-term potential population of user u is characterized, respectively. G v Representing a potential group of users. b s u and bl u Representing probability distributions for the short-term and long-term user groups, respectively, to which the user is currently likely to belong.
Step 5: constructing a hybrid user characterization by using the long-term session characterization and the short-term session characterization of the user obtained in the step 2 and the group influence characterization obtained in the step 4; the step 5 specifically comprises the following steps: constructing a mixed user representation by utilizing the user long-short term session representation obtained in the step 2 and the influence representation of the current population and the history population obtained in the step 4, and carrying out { x } s u ,x l u ,g s u ,g l u Denoted as F;
combining x in a dynamic manner s u 、x l u 、g s u 、g l u Specific methods are using MLP polymerization or using attention polymerization;
polymerization using MLP: each feature vector is mapped to a scalar using a multi-layer perceptron, and then the scalar is converted to the weight of each component using a softmax layer, as specifically calculated below:
where F represents a blended user representation, F represents a feature vector in F, w f Weights, h, generated for feature vector f u Representing the resulting hybrid user characterization. MLP denotes a multi-layer perceptron, T denotes a transpose operation, exp () denotes an exponential function based on e.
Attention aggregation was used: the weight is calculated according to the relation between the component and the representation of the target user u, and the concrete calculation is as follows:
wherein F represents a mixed user representation, F represents a certain feature vector in F, u represents a static representation of user u, w f Weights, h, generated for feature vector f u Representing the resulting hybrid user characterization. T denotes a transpose operation, exp () denotes an exponential function based on e.
Step 6: estimating the probability of one item becoming the next access item by using the mixed user characterization obtained in the step 5 to replace the user potential vector in the paired model; the step 6 is specifically as follows: calculating step 5 to obtain the hybrid user representation h u The inner product of the product and the characterization vector of the item v is taken as the preference score of the predicted user u on the item v, and the calculation formula is as follows:
wherein Representing a predicted preference score, h, of user u for item v u Representing a hybrid user characterization, v representing a static characterization of the item v, and T representing a transpose operation.
Training a model with ordering and pairwise penalty functions for positive samples v + Randomly selecting an item from the user's current session, for negative sampling v - Selecting a commodity that a user has never purchased or accessed before, the final loss function is as follows:
wherein and />Representing the predicted user u versus the sample item v, respectively + And negatively sampled item v - Preference score of->Representing a long-term user session. Sigma represents the s igmoid function, θ represents a parameter of the model, argmin represents the minimization function.
And selecting the item with the highest prediction score as the item which is most likely to be accessed next by the predicted user u by using the obtained trained model.
The method of the embodiment is a recommendation method based on long-term and short-term potential user group modeling, and the user characterization with more abundant information is learned and obtained by considering the long-term and short-term session of the user and different influences of the potential user group on the current preference of the modeling user, so that the next item which should be recommended to the user is predicted more accurately.
The method of the embodiment has practical significance and good application prospect in the aspect of modeling the dynamic preference of the user by considering the evolution of the user group interest information.
Since the behavior patterns of users are complex, in order to capture the dynamic and evolving groupings of each potential user group, the problem to be solved by the method of the present embodiment is manifested by the following points:
a. in a real scenario, a user may belong to multiple user groups, rather than being assigned a static group as in the traditional local approach;
b. the interests of each potential user population will be updated over time;
c. users may switch their potential groups for a variety of reasons, such as evolution of user preferences and demands.
Example 2:
the embodiment provides a user characterization learning and recommending system, which comprises the following steps: :
module M1: using a fully connected neural network to convert one-hot codes of users and projects into an ebedding representation;
module M2: dividing a user session into a long-term session and a short-term session, extracting key information based on the representation of the items in the session obtained in the module M1, and respectively forming the long-term session representation and the short-term session representation through a pooling layer;
module M3: based on the long-term and short-term session characterization obtained by the module M2, calculating probability distribution of a user group to which the user possibly belongs;
module M4: according to the probability distribution obtained by the module M3, the characteristics of different user groups are aggregated, the influence of the user neighbors and the preference difference between the aspiration channel user subsets are captured, and the group influence characterization is obtained;
module M5: constructing a hybrid user characterization by using the long-term session characterization and the short-term session characterization of the user obtained by the module M2 and the group influence characterization obtained by the module M4;
module M6: the mixed user token obtained by module M5 is used to estimate the probability of one item becoming the next access item instead of the user potential vector in the paired model.
Example 3:
the present embodiment will be understood by those skilled in the art as more specific descriptions of embodiment 1 and embodiment 2.
The embodiment provides a recommendation method based on long-term and short-term potential user group modeling, which comprises the following steps:
step 1: using a fully connected neural network to convert one-hot codes of users and articles into an ebedding representation;
step 2: dividing a user session into a long-term session and a short-term session, extracting key information based on the representation of the items in the session, and respectively forming the long-term session representation and the short-term session representation through a pooling layer:
step 3: based on the long-short session characterization, calculating probability distribution of a user group to which the user possibly belongs;
step 4: aggregating features of different user groups to capture the influence of user neighbors and preference differences between subsets of like-minded users;
step 5: constructing a hybrid user characterization using the long-term and short-term session characterizations of the user and the group influence;
step 6: the resulting hybrid user token is used to replace the user potential vector in the pairwise model to estimate the probability of one item being the next access item.
In order to consider the influence of potential user groups to better understand the user preference, the embodiment proposes to learn the user characterization by modeling long-term and short-term dynamic potential user groups, the model learns the long-term and short-term session characterization of the user respectively by using two networks, then uses two other units to detect which potential user groups the user belongs to, and carries out weighted aggregation on the characterization of the potential groups, then comprehensively considers the four aspects based on the attention mechanism to form the final user preference characterization, finally uses the characterization to replace the potential vector of the user in the paired model to estimate the probability that one item becomes the next access item, and the model can capture the potential group information of the user and can also sense the change condition of the user group information with time.
Example 4:
the present embodiment will be understood by those skilled in the art as more specific descriptions of embodiment 1 and embodiment 2.
As shown in tables 1 and 2, assume that the user set and the item set are respectivelyAndfor one user u in the set of users, let +.>Is the sequence of items it clicks on in one session (e.g., Δt time). By analysing the whole behavioral history of user u, his session sequence can be obtained>Where t represents the session index corresponding to the timestamp.
Table 1 data set statistics table
Table 2 problem definition symbol descriptive table
Assuming there are L potential user groups, matrix G v ∈R L×K The preferences of these user groups are described, together with matrix G k ∈R L×K DeterminesRelationship of users and potential user groups. Whereby a user can be assigned to a plurality of groups and the cumulative impact of the group of users on that user can be obtained.
Thus, given a user u and his session sequenceBy taking into account the long-term and short-term sessions of user u, and the influence of his associated long-term and short-term potential population, the next most likely to belong to the current session s is predicted t Is a project of (2).
The embodiment provides a recommendation method based on long-term and short-term potential user group modeling, which comprises the following specific steps:
step 1: using a fully connected neural network to convert one-hot codes of users and articles into ebedding representation to obtain a user matrix U E R N×K And an item matrix V E R M×K Wherein n= |u|, m= |v| represent the number of users and items, respectively, K represents the implicit vector dimension, U e R K and v∈RK Representing static representations of user u and item v, respectively;
step 2: dividing a session of user u into two parts, wherein the current sessionFor short-term sessions, the remaining sessions are used to describe the short-term needs of the userFor long-term sessions, for representing the user's long-term preferences, generating an item representation matrix for each session based on the item representations in that session, and aggregating the item features through a pooling operation to form two context-aware input representations x s and xl Reflecting the state of the user without the time stamp, the specific calculation formula is as follows:
wherein ,xs and xl A short-term session characterization and a long-term session characterization respectively,representing a matrix spliced by static representations of items in a short-term session, < >>Representing a matrix stitched from static representations of items in a long-term session. pooling represents a weighted average pooling operation. According to the static representation of the target user u and the static representation of the object v, calculating the weight according to the specific calculation formula as follows:
wherein ,wu,v Representing the weight generated for item v based on target user u, u representing the static characterization of user u, v representing the static characterization of item v,representing a long-term user session, v i Representing a static characterization of the ith item in the long-term session. T denotes a transpose operation, exp () denotes an exponential function based on e.
Step 3: long-short session characterization x generated based on step 2 s and xl (input representation) computing them with potential user group G k Similarity between key representations of (c) to evaluate probability distribution b of a group of users to which the user is currently likely to belong s u and bs u The calculation formula is as follows:
wherein ,bs u ,b l u ∈R L Representing probability distribution, x, of short-term and long-term user groups, respectively, to which a user is currently likely to belong s and xl Representing short-term session characterization and long-term session characterization, respectively. The softmax function will vector G k x s u And converting into a pseudo probability distribution vector.
Step 4: according to the probability distribution b obtained in step 3 s u and bs u Features of different user groups are aggregated to capture the impact of user neighbors and preference differences between subsets of like-minded users. The calculation formula of the aggregate user group characteristics is as follows:
wherein ,gl u ∈R K ,g s u ∈R K The impact of the long-term and short-term potential population of user u is characterized, respectively. G v Representing a potential group of users. b s u and bl u Representing probability distributions for the short-term and long-term user groups, respectively, to which the user is currently likely to belong.
Step 5: constructing a mixed user characterization by using the user long-short term session characterization (personal preference characterization and the influence characterization (group influence characterization) of the current group and the history group obtained in the step 2,will { x } s u ,x l u ,g s u ,g l u Denoted as F, to combine these four components in a dynamic manner, the present embodiment provides two methods to fuse them:
1) Polymerization using MLP: each feature vector is mapped to a scalar using a multi-layer perceptron (MLP), and then the scalar is converted to the weight of each component using a softmax layer, as specifically calculated below:
where F represents a blended user representation, F represents a feature vector in F, w f Weights, h, generated for feature vector f u Representing the resulting hybrid user characterization. MLP denotes a multi-layer perceptron, T denotes a transpose operation, exp () denotes an exponential function based on e.
2) Attention aggregation was used: the weight is calculated according to the relation between the component and the representation of the target user u, and the concrete calculation is as follows:
wherein F represents a mixed user representation, F represents a certain feature vector in F, u represents a static representation of user u, w f Weights, h, generated for feature vector f u Representing the resulting hybrid user characterization. T denotes a transpose operation, exp () denotes an exponential function based on e.
Step 6: obtaining a hybrid user representation h in step 5 u Then, the inner product of the product and the characterization vector of the item v is calculated, and the product is taken as the predicted preference score of the user u on the item v, wherein the calculation formula is as follows:
wherein Representing a predicted preference score, h, of user u for item v u Representing a hybrid user characterization, v representing a static characterization of the item v, and T representing a transpose operation.
Step 7: training a model with ordering and pairwise penalty functions for positive samples v + Randomly selecting an item from the user's current session, for negative sampling v - A commodity that a user has never purchased or accessed before is selected and the final loss function is as follows:
wherein and />Representing the predicted user u versus the sample item v, respectively + And negatively sampled item v - Preference score of->Representing a long-term user session. Sigma represents the s igmoid function, θ represents a parameter of the model, argmin represents the minimization function.
Step 8: and (3) selecting the item with the highest prediction score as the item which is most likely to be accessed next by the predicted user u by using the trained model obtained in the step (7).
The invention solves the problem that in most of the existing session-based recommendation systems, the characterization of the user is summarized based on the session of the user independently, and the information sharing between user models is ignored.
Those skilled in the art will appreciate that the invention provides a system and its individual devices, modules, units, etc. that can be implemented entirely by logic programming of method steps, in addition to being implemented as pure computer readable program code, in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers, etc. Therefore, the system and various devices, modules and units thereof provided by the invention can be regarded as a hardware component, and the devices, modules and units for realizing various functions included in the system can also be regarded as structures in the hardware component; means, modules, and units for implementing the various functions may also be considered as either software modules for implementing the methods or structures within hardware components.
The foregoing describes specific embodiments of the present invention. It is to be understood that the invention is not limited to the particular embodiments described above, and that various changes or modifications may be made by those skilled in the art within the scope of the appended claims without affecting the spirit of the invention. The embodiments of the present application and features in the embodiments may be combined with each other arbitrarily without conflict.
Claims (10)
1. The user characterization learning and recommending method is characterized by comprising the following steps:
step 1: using a fully connected neural network to convert one-hot codes of users and projects into an ebedding representation;
step 2: dividing a user session into a long-term session and a short-term session, extracting key information based on the characterization of the items in the session obtained in the step 1, and respectively forming the long-term session characterization and the short-term session characterization through a pooling layer;
step 3: calculating probability distribution of a user group to which the user possibly belongs based on the long-term and short-term session characterization obtained in the step 2;
step 4: according to the probability distribution obtained in the step 3, the characteristics of different user groups are aggregated, the influence of the user neighbors and the preference difference between the aspiration channel user subsets are captured, and the group influence characterization is obtained;
step 5: constructing a hybrid user characterization by using the long-term session characterization and the short-term session characterization of the user obtained in the step 2 and the group influence characterization obtained in the step 4;
step 6: and (3) using the mixed user characterization obtained in the step 5 to replace the user potential vector in the paired model, and estimating the probability of one item to be the next access item.
2. The method for learning and recommending user features according to claim 1, wherein the step 1 specifically comprises: using a fully connected neural network to convert one-hot codes of users and articles into low-dimensional ebedding representation to obtain a user matrix U epsilon R N×K And an item matrix V E R M×K Wherein n= |u|, m= |v| represent the number of users and the number of items, respectively, K represents the implicit vector dimension, U e R K and v∈RK Representing the static representation of user u and the static representation of item v, respectively, R representing the matrix.
3. The method for learning and recommending user features according to claim 2, wherein the step 2 specifically comprises: dividing a session of a user u into two parts, wherein the current sessionFor short-term session, the user's short-term needs are described, the remaining sessions are +.>For long-term session, for representing long-term preference of user, wherein v i Representing a static representation of the ith item, +.>Representing the number of items in a short-term session, +.>Representing the number of items in the long-term session;
based on the item representations in each session, an item representation matrix is generated for the session and the item features are aggregated by a pooling operation to form two context-aware input representations x s and xl Reflecting the user's status without a time stamp.
4. The user characterization learning and recommendation method according to claim 3, wherein the specific calculation formula is as follows:
wherein ,xs and xl A short-term session characterization and a long-term session characterization respectively,representing a matrix spliced by static representations of items in a short-term session, < >>Representing a matrix stitched from static representations of items in a long-term session. The pooling represents a weighted average pooling operation, and the weight is calculated according to the static representation of the target user u and the static representation of the object v, and the specific calculation formula is as follows:
wherein ,wu,v Representing the weight generated for item v based on target user u, u representing the static characterization of user u, v representing the static characterization of item v,representing a long-term user session, v i Representing a static representation of the ith item in the long-term session, T represents a transpose operation, exp () represents an exponential function based on e.
5. The method for learning and recommending user features according to claim 4, wherein the step 3 specifically comprises: long-short session characterization x generated based on step 2 s and xl Calculate them and potential user group G k Similarity between key representations of (a) and (b) evaluating probability distribution b of a group of users to which a user is currently likely to belong s u and bl u The specific calculation formula is as follows:
wherein ,bs u ,b l u ∈R L Representing probability distribution, x, of short-term and long-term user groups, respectively, to which a user is currently likely to belong s and xl Representing short-term session characterization and long-term session characterization, respectively, the softmax function will vector G k x s u And converting into a pseudo probability distribution vector.
6.The method for learning and recommending user features according to claim 5, wherein the step 4 specifically comprises: according to the probability distribution b obtained in step 3 s u and bl u The characteristics of different user groups are aggregated, the influence of the neighbors of the users and the preference difference among the subsets of the like-minded users are captured, and the calculation formula of the characteristics of the aggregated user groups is as follows:
wherein ,gl u ∈R K ,g s u ∈R K Characterizing the influence of the long-term and short-term potential population of user u, G v Representing potential user groups, b s u and bl u Representing probability distributions for the short-term and long-term user groups, respectively, to which the user is currently likely to belong.
7. The method for learning and recommending user features according to claim 6, wherein the step 5 specifically comprises: constructing a mixed user representation by utilizing the user long-short term session representation obtained in the step 2 and the influence representation of the current population and the history population obtained in the step 4, and carrying out { x } s u ,x l u ,g s u ,g l u Denoted as F.
8. The method of claim 7, wherein x is combined in a dynamic manner s u 、x l u 、g s u 、g l u Specific methods are using MLP polymerization or using attention polymerization;
polymerization using MLP: each feature vector is mapped to a scalar using a multi-layer perceptron, and then the scalar is converted to the weight of each component using a softmax layer, as specifically calculated below:
wherein F represents a mixed user representation, F represents a certain feature vector in F, w f Weights, h, generated for feature vector f u Representing the resulting hybrid user representation, MLP representing the multi-layer perceptron, T representing the transpose operation, exp () representing the e-based exponential function;
attention aggregation was used: the weight is calculated according to the relation between the component and the representation of the target user u, and the concrete calculation is as follows:
wherein F represents a mixed user representation, F represents a certain feature vector in F, u represents a static representation of user u, and w f Weights, h, generated for feature vector f u Representing the resulting hybrid user token, T represents the transpose operation and exp () represents the base e exponential function.
9. The method for learning and recommending user features according to claim 8, wherein the step 6 specifically comprises: calculating step 5 to obtain the hybrid user representation h u Inner product with the token vector of item v as a predictive user u preference for item vThe number and the calculation formula are as follows:
wherein ,representing a predicted preference score, h, of user u for item v u Representing a hybrid user characterization, v representing a static characterization of the item v, T representing a transpose operation;
training a model with ordering and pairwise penalty functions for positive samples v + Randomly selecting an item from the user's current session, for negative sampling v - Selecting a commodity that a user has never purchased or accessed before, the final loss function is as follows:
wherein , and />Representing the predicted user u versus the sample item v, respectively + And negatively sampled item v - Is a function of the preference score of (c),representing a long-term user session, σ representing a sigmoid function, θ representing parameters of the model, argmin representing a minimization function;
and selecting the item with the highest prediction score as the item which is most likely to be accessed next by the predicted user u by using the obtained trained model.
10. A user characterization learning and recommendation system, comprising the steps of:
module M1: using a fully connected neural network to convert one-hot codes of users and projects into an ebedding representation;
module M2: dividing a user session into a long-term session and a short-term session, extracting key information based on the representation of the items in the session obtained in the module M1, and respectively forming the long-term session representation and the short-term session representation through a pooling layer;
module M3: based on the long-term and short-term session characterization obtained by the module M2, calculating probability distribution of a user group to which the user possibly belongs;
module M4: according to the probability distribution obtained by the module M3, the characteristics of different user groups are aggregated, the influence of the user neighbors and the preference difference between the aspiration channel user subsets are captured, and the group influence characterization is obtained;
module M5: constructing a hybrid user characterization by using the long-term session characterization and the short-term session characterization of the user obtained by the module M2 and the group influence characterization obtained by the module M4;
module M6: the mixed user token obtained by module M5 is used to estimate the probability of one item becoming the next access item instead of the user potential vector in the paired model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310135479.9A CN116258551A (en) | 2023-02-17 | 2023-02-17 | User characterization learning and recommending method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310135479.9A CN116258551A (en) | 2023-02-17 | 2023-02-17 | User characterization learning and recommending method and system |
Publications (1)
Publication Number | Publication Date |
---|---|
CN116258551A true CN116258551A (en) | 2023-06-13 |
Family
ID=86683902
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310135479.9A Pending CN116258551A (en) | 2023-02-17 | 2023-02-17 | User characterization learning and recommending method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116258551A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117094360A (en) * | 2023-10-18 | 2023-11-21 | 杭州同花顺数据开发有限公司 | User characterization extraction method, device, equipment and storage medium |
-
2023
- 2023-02-17 CN CN202310135479.9A patent/CN116258551A/en active Pending
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117094360A (en) * | 2023-10-18 | 2023-11-21 | 杭州同花顺数据开发有限公司 | User characterization extraction method, device, equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111177575B (en) | Content recommendation method and device, electronic equipment and storage medium | |
Chen et al. | Sequence-aware factorization machines for temporal predictive analytics | |
CN111797321B (en) | Personalized knowledge recommendation method and system for different scenes | |
CN110503531B (en) | Dynamic social scene recommendation method based on time sequence perception | |
Lee et al. | Scalable deep learning-based recommendation systems | |
Yassine et al. | Intelligent recommender system based on unsupervised machine learning and demographic attributes | |
CN109885756B (en) | CNN and RNN-based serialization recommendation method | |
Lin et al. | A survey on reinforcement learning for recommender systems | |
Wen et al. | Neural attention model for recommendation based on factorization machines | |
CN112381581A (en) | Advertisement click rate estimation method based on improved Transformer | |
CN111950593A (en) | Method and device for recommending model training | |
Navgaran et al. | Evolutionary based matrix factorization method for collaborative filtering systems | |
CN111222039A (en) | Session recommendation method and system based on long-term and short-term interest combination | |
Zhao et al. | Usersim: User simulation via supervised generativeadversarial network | |
WO2024041483A1 (en) | Recommendation method and related device | |
CN116258551A (en) | User characterization learning and recommending method and system | |
Hassan et al. | Performance analysis of neural networks-based multi-criteria recommender systems | |
Luo et al. | BGNN: Behavior-aware graph neural network for heterogeneous session-based recommendation | |
Wang et al. | The application of factorization machines in user behavior prediction | |
Nazari et al. | Scalable and data-independent multi-agent recommender system using social networks analysis | |
CN116910357A (en) | Data processing method and related device | |
Chen et al. | Attention-enhanced graph neural networks with global context for session-based recommendation | |
Long et al. | Hierarchical attention factorization machine for CTR prediction | |
Kumar et al. | Session-based recommendations with sequential context using attention-driven LSTM | |
Yang et al. | Collaborative meta-path modeling for explainable recommendation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |